Faster-ILOD

This project hosts the code for implementing the Faster ILOD algorithm for incremental object detection, as presented in our paper:

Faster ILOD: Incremental Learning for Object Detectors based on Faster RCNN

Can Peng, Kun Zhao and Brian C. Lovell; In: Pattern Recognition Letters 2020.

arXiv preprint.

Installation

This Faster ILOD implementation is based on maskrcnn-benchmark. Therefore the installation is the same as the original maskrcnn-benchmark.

Please check INSTALL.md for installation instructions. You may also want to see the original README.md of maskrcnn-benchmark.

Training

The files used to train Faster ILOD models are under Faster-ILOD/tools folder.

train_first_step.py: normally train the first task (standard training).

train_incremental.py: incrementally train the following tasks (knowledge distillation based training).

The config settings for the models and datasets are under Faster-ILOD/configs folder.

VOC dataset training

e2e_faster_rcnn_R_50_C4_1x_Source_model.yaml: config and dataset settings for source model (ResNet50) trained on VOC dataset.

e2e_faster_rcnn_R_50_C4_1x_Target_model.yaml: config and dataset settings for target model (ResNet50) trained on VOC dataset.

The code for loading VOC dataset to the model is written on the file Faster-ILOD/maskrcnn_benchmark/data/datasets/voc.py.

1. Please modify the path for putting VOC dataset on the file Faster-ILOD/maskrcnn_benchmark/config/paths_catalog.py.

2. Please modify the setting for the name of old class categories (all previously trained categories) on NAME_OLD_CLASSES on the file e2e_faster_rcnn_R_50_C4_1x_Target_model.yaml.

3. Please modify the setting for the name of new class categories (categories for current training task) on NAME_NEW_CLASSES on the file e2e_faster_rcnn_R_50_C4_1x_Target_model.yaml.

4. Please modify the setting for the name of excluded categories (categories not used, since VOC has 20 categories) on NAME_EXCLUDED_CLASSES on the file e2e_faster_rcnn_R_50_C4_1x_Target_model.yaml.

5. Please modify the number of detecting categories on NUM_CLASSES on the file e2e_faster_rcnn_R_50_C4_1x_Source_model.yaml (number of old categories) and the file e2e_faster_rcnn_R_50_C4_1x_Target_model.yaml (number of old and new categories), repectively.

COCO dataset training

e2e_faster_rcnn_R_50_C4_1x_Source_model_COCO.yaml: config and dataset settings for source model (ResNet50) trained on COCO dataset.

e2e_faster_rcnn_R_50_C4_1x_Target_model_COCO.yaml: config and dataset settings for target model (ResNet50) trained on COCO dataset.

The code for loading COCO dataset to the model is written on the file Faster-ILOD/maskrcnn_benchmark/data/datasets/coco.py.

1. Please modify the path for putting COCO dataset on the file Faster-ILOD/maskrcnn_benchmark/config/paths_catalog.py.

2. The categories for COCO dataset training are added in alphabetical orders. Please modify the number of detecting categories on NUM_CLASSES on the file e2e_faster_rcnn_R_50_C4_1x_Source_model_COCO.yaml (number of old categories) and the file e2e_faster_rcnn_R_50_C4_1x_Target_model_COCO.yaml (number of old and new categories), repectively.

Distillation Loss

The code for calculating feature, RPN, and RCN distillation losses are written on the file Faster-ILOD/blob/main/maskrcnn_benchmark/distillation/distillation.py.

Citations

Please consider citing the following paper in your publications if it helps your research.

@article{peng2020faster,
  title={Faster ILOD: Incremental Learning for Object Detectors based on Faster RCNN},  
  author={Peng, Can and Zhao, Kun and Lovell, Brian C},  
  journal={Pattern Recognition Letters},  
  year={2020} 
}

Acknowledgements

Our Faster ILOD implementation is based on maskrcnn-benchmark. We thanks the authors for making their code public.